README.md

metadata

license: mit
tags:
  - yacht
  - agent
  - jax
  - flax
  - RL
pipeline_tag: reinforcement-learning

For detailed explanations and dev logs, please check out the Github Repository!

Models

Yacht Agent: REINFORCE

• Model Architecture

  • Input: Gymnax Observation space -> 19 dimensions
  • Hidden1: 128 units, Leaky ReLU
  • Hidden2: 64 units, Leaky ReLU
  • Hidden3: 64 units, Leaky ReLU
  • Output: Action space -> 44 dimensions

• Training: RTX 4070TI SUPER * 3 hours -> 8B Games

Environment

• Python 3.11
• Main Framework: JAX, FLAX, gymnax

About Yacht

The Yacht game in this Yacht-Agent follows the rules of Nintendo's Clubhouse Games: 51 Worldwide Classics. Below is a brief description.

1. The object of the game is to score points by rolling five dice to make certain combinations.
2. The dice can be rolled up to three times in a turn to try to make these combinations.
3. After each round the player chooses which scoring category is to be used for that round.
4. Once a category has been used in the game, it cannot be used again.
5. The winner is the player who scores the most points.

Yacht Scoring

Category Condition Score Aces Any combination Sum of dice with number 1 Deuces Any combination Sum of dice with number 2 Threes Any combination Sum of dice with number 3 Fours Any combination Sum of dice with number 4 Fives Any combination Sum of dice with number 5 Sixes Any combination Sum of dice with number 6 (Bonus) Score of Aces...Sixes >= 63 +35 points Choice Any combination Sum of all dice 4 of a Kind At least 4 dice the same Sum of all dice Full House 3 of one number, 2 of another Sum of all dice S. Straight 4 sequential dice 15 points L. Straight 5 sequential dice 30 points Yacht All 5 dice the same 50 points